This research is a basic study on utilizing artificial intelligence (AI) by applying deep learning to the fields of architecture and urban design.
In recent years, the use of budding technologies, such as deep learning, has increased in the field of architecture and urban design. While this technology has potential in various fields, this study focuses on learning and deduction of sensibility evaluation and impression of design. Needless to say, the relationship of design with sensibility and impression is important, and the design heightens sensibility and impressions. However, the causal relationship of quantitative representation (feature value) and feature value of design with impression is complex and is characteristically difficult to deduce. Such a characteristic is a property similar to fields where deep learning has been successfully used. It is, therefore, thought that AI using deep learning could be applicable. As mentioned, this research examines the budding properties of AI that deduce “street names and desire to visit” based on city landscapes. Specifically, the “desire/no-desire to visit (classification)” and “degree of desire to visit” are deduced, and as constituents of image consciousness, street names are also classified (21 classes).
The object of the study and the city landscapes were prepared from 21 streets selected from a large city and sightseeing information. The images for city landscapes were obtained from street view on Google Earth to ensure that these images were not of any one building, ground, or sky. A total of 2,100 images, 100 for each street, were considered.
Deduction AI with high precision was first successfully developed to deduce “classification of street names (21 classes)”. Its precision was approximately 86% for the F-value with a K-coefficient of 0.8508 (p-value = 1.6e-15). Next, for the classification, deduction AI with high conformity with desire/no-desire to visit criteria of test subjects was successfully prepared. Its precision had a K-coefficient of 0.8920 (p-value = 2.2e-15). Further, for deducing degree of desire, there was little difference in the degree of desire to visit between test subjects, and AI permitting deduction with high correlation was successfully developed. For its precision, the effect size of Wilcoxon’s signed rank test (test of paired nonparametric data) was 0.18, and Spearman’s rank correlation was 0.7564 (p-value = 0.0005742). Finally, to generalize the methodology of AI using deep learning, the 95% confidence interval that considered 100 kinds of AI developed using this method was confirmed to be small. Specifically, the effect size did not exceed 0.2 (a threshold value indicating small effect size) and did not fall below 0.6 (a threshold value indicating high correlation). Under the experimental conditions of this study, the AI developed using deep learning can be described as a method that presents generality in the degree of precision.
From the perspective of the experimental conditions of the study and usage, a successful impression deduction AI for city landscapes with good precision is developed. This provides the first step in systematically organizing and investigating the hitherto unstudied budding potential of deep learning in the fields of architecture and urban design.